Thank you for working on this! There's a few things that I think should be changed, let me know what you think:

First, I feel like loading from just one folder with fixed names is less than clear. The CLIP/mT5 model being separately selectable would make more sense, then we can split it up like this:

• chinese-roberta-wwm-ext-large.bin goes in models/clip (technically not a CLIP model, not sure why they named it that?)
• mt5-xl.bin goes in models/t5
• HunyuanDiT.pt goes in models/unet

(the user would have to rename the files from just "pytorch_model.bin")

Second, there's a lot of small changes that need to be made, and some code redundancy that needs removed. For example, the MT5/CLIP are currently non-standard, and don't work with the regular text encode node, though this might be hard to implement.

I can't really merge this as-is, but I can adapt/fix parts of it and add it to the main repo if that works for you. I'll add your name to the readme for credit.

Oh, and I noticed you're saving the embeddings for testing. Could you send me those for easy debugging? clip_prompt_embeds.pt, etc. It would help a lot.

I agree with what you said. I have added support for custom resolution, batch image generation, and img2img generation. If there is anything worth referring to, you can go directly https://github.com/chaojie/ComfyUI_ExtraModels

I've got the initial proof of concept working (trying to integrate it more closely). Custom resolutions/etc will still need more work and it's a WIP at the moment, with a lot of it using the original reference code.

Again, thank you for your work, I will continue to reference it as I fix things.